How do you organise a python project that contains multiple packages so that each file in a package can still be run individually?

Question

TL;DR

Here's an example repository that is set up as described in the first diagram (below): https://github.com/Poddster/package_problems

If you could please make it look like the second diagram in terms of project organisation and can still run the following commands, then you've answered the question:

$ git clone https://github.com/Poddster/package_problems.git
$ cd package_problems
<do your magic here>

$ nosetests

$ ./my_tool/my_tool.py
$ ./my_tool/t.py
$ ./my_tool/d.py

 (or for the above commands, $ cd ./my_tool/ && ./my_tool.py is also acceptable)

Alternatively: Give me a different project structure that allows me to group together related files ('package'), run all of the files individually, import the files into other files in the same package, and import the packages/files into other package's files.

---

Current situation

I have a bunch of python files. Most of them are useful when callable from the command line i.e. they all use argparse and if __name__ == "__main__" to do useful things.

Currently I have this directory structure, and everything is working fine:

.
├── config.txt
├── docs/
│   ├── ...
├── my_tool.py
├── a.py
├── b.py
├── c.py
├── d.py
├── e.py
├── README.md
├── tests
│   ├── __init__.py
│   ├── a.py
│   ├── b.py
│   ├── c.py
│   ├── d.py
│   └── e.py
└── resources
    ├── ...

Some of the scripts import things from other scripts to do their work. But no script is merely a library, they are all invokable. e.g. I could invoke ./my_tool.py, ./a.by, ./b.py, ./c.py etc and they would do useful things for the user.

"my_tool.py" is the main script that leverages all of the other scripts.

What I want to happen

However I want to change the way the project is organised. The project itself represents an entire program useable by the user, and will be distributed as such, but I know that parts of it will be useful in different projects later so I want to try and encapsulate the current files into a package. In the immediate future I will also add other packages to this same project.

To facilitate this I've decided to re-organise the project to something like the following:

.
├── config.txt
├── docs/
│   ├── ...
├── my_tool
│   ├── __init__.py
│   ├── my_tool.py
│   ├── a.py
│   ├── b.py
│   ├── c.py
│   ├── d.py
│   ├── e.py
│   └── tests
│       ├── __init__.py
│       ├── a.py
│       ├── b.py
│       ├── c.py
│       ├── d.py
│       └── e.py
├── package2
│   ├── __init__.py
│   ├── my_second_package.py
|   ├── ...
├── README.md
└── resources
    ├── ...

However, I can't figure out an project organisation that satisfies the following criteria:

1. All of the scripts are invokable on the command line (either as my_tool\a.py or cd my_tool && a.py)
2. The tests actually run :)
3. Files in package2 can do import my_tool

The main problem is with the import statements used by the packages and the tests.

Currently, all of the packages, including the tests, simply do import <module> and it's resolved correctly. But when jiggering things around it doesn't work.

Note that supporting py2.7 is a requirement so all of the files have from __future__ import absolute_import, ... at the top.

What I've tried, and the disastrous results

1

If I move the files around as shown above, but leave all of the import statements as they currently are:

1. $ ./my_tool/*.py works and they all run properly
2. $ nosetests run from the top directory doesn't work. The tests fail to import the packages scripts.
3. pycharm highlights import statements in red when editing those files :(

2

If I then change the test scripts to do:

from my_tool import x

1. $ ./my_tool/*.py still works and they all run properly
2. $ nosetests run from the top directory doesn't work. Then tests can import the correct scripts, but the imports in the scripts themselves fail when the test scripts import them.
3. pycharm highlights import statements in red in the main scripts still :(

3

If I keep the same structure and change everything to be from my_tool import then:

1. $ ./my_tool/*.py results in ImportErrors
2. $ nosetests runs everything ok.
3. pycharm doesn't complain about anything

e.g. of 1.:

Traceback (most recent call last):
  File "./my_tool/a.py", line 34, in <module>
    from my_tool import b
ImportError: cannot import name b

4

I also tried from . import x but that just ends up with ValueError: Attempted relative import in non-package for the direct running of scripts.

Looking at some other SO answers:

I can't just use python -m pkg.tests.core_test as

a) I don't have main.py. I guess I could have one?
b) I want to be able to run all of the scripts, not just main?

I've tried:

if __name__ == '__main__' and __package__ is None:
    from os import sys, path
    sys.path.append(path.dirname(path.dirname(path.abspath(__file__))))

but it didn't help.

I also tried:

__package__ = "my_tool"
from . import b

But received:

SystemError: Parent module 'loading_tool' not loaded, cannot perform relative import

adding import my_tool before from . import b just ends up back with ImportError: cannot import name b

Fix?

What's the correct set of magical incantations and directory layout to make all of this work?

Solution 1

Once you move to your desired configuration, the absolute imports you are using to load the modules that are specific to my_tool no longer work.

You need three modifications after you create the my_tool subdirectory and move the files into it:

1. Create my_tool/__init__.py. (You seem to already do this but I wanted to mention it for completeness.)

2. In the files directly under in my_tool: change the import statements to load the modules from the current package. So in my_tool.py change:

   import c
   import d
   import k
   import s
   

   to:

   from . import c
   from . import d
   from . import k
   from . import s
   

   You need to make a similar change to all your other files. (You mention having tried setting __package__ and then doing a relative import but setting __package__ is not needed.)

3. In the files located in my_tool/tests: change the import statements that import the code you want to test to relative imports that load from one package up in the hierarchy. So in test_my_tool.py change:

   import my_tool
   

   to:

   from .. import my_tool
   

   Similarly for all the other test files.

With the modifications above, I can run modules directly:

$ python -m my_tool.my_tool
C!
D!
F!
V!
K!
T!
S!
my_tool!
my_tool main!
|main tool!||detected||tar edit!||installed||keys||LOL||ssl connect||parse ASN.1||config|

$ python -m my_tool.k
F!
V!
K!
K main!
|keys||LOL||ssl connect||parse ASN.1|

and I can run tests:

$ nosetests 
........
----------------------------------------------------------------------
Ran 8 tests in 0.006s

OK

Note that I can run the above both with Python 2.7 and Python 3.

---

Rather than make the various modules under my_tool be directly executable, I suggest using a proper setup.py file to declare entry points and let setup.py create these entry points when the package is installed. Since you intend to distribute this code, you should use a setup.py to formally package it anyway.

1. Modify the modules that can be invoked from the command line so that, taking my_tool/my_tool.py as example, instead of this:

   if __name__ == "__main__":
       print("my_tool main!")
       print(do_something())
   

   You have:

   def main():
       print("my_tool main!")
       print(do_something())
   
   if __name__ == "__main__":
       main()
   

2. Create a setup.py file that contains the proper entry_points. For instance:

   from setuptools import setup, find_packages
   
   setup(
       name="my_tool",
       version="0.1.0",
       packages=find_packages(),
       entry_points={
           'console_scripts': [
               'my_tool = my_tool.my_tool:main'
           ],
       },
       author="",
       author_email="",
       description="Does stuff.",
       license="MIT",
       keywords=[],
       url="",
       classifiers=[
       ],
   )
   

   The file above instructs setup.py to create a script named my_tool that will invoke the main method in the module my_tool.my_tool. On my system, once the package is installed, there is a script located at /usr/local/bin/my_tool that invokes the main method in my_tool.my_tool. It produces the same output as running python -m my_tool.my_tool, which I've shown above.

Solution 2

Point 1

I believe it's working, so I don't comment on it.

Point 2

I always used tests at the same level as my_tool, not below it, but they should work if you do this at the top of each tests files (before importing my_tool or any other py file in the same directory)

import os
import sys

sys.path.insert(0, os.path.abspath(__file__).rsplit(os.sep, 2)[0])

Point 3

In my_second_package.py do this at the top (before importing my_tool)

import os
import sys

sys.path.insert(0,
                os.path.abspath(__file__).rsplit(os.sep, 2)[0] + os.sep
                + 'my_tool')

Best regards,

JM

Comments

[Pod]

Thanks, this fits all the requirements. I'm not keen on invoking it as python -m my_tool.my_tool as it's not shell-friendly (i.e. no tab-complete, no obeying +x permission). You didn't mention it, but it also works with other packages in the same project loading my_tool, assuming they are also envoked with python -m my_second_package.blah. Do you have any suggestions for a more shell-friendly versions? If not, I might borrow a bit from @CasualDemon's answer, i.e. catch ImportError for when they're directly invoked, or just make a front-end script that takes the script to run as an arg.

[Pod]

It works, though pycharm still thinks the modules don't exist, even after tinkering with it's project structure config. Note: I still intended to import my_tool into my_second_package via import my_tool.whatever rather than simply import whatever, but all I had to do is not add "/my_tool" to the path, so point3 shares the same code as point 2.

[Pod]

Not sure what the downvote was for: It works against the criteria I asked for, including being invoked as tool/blah.py. You even worked with the test repo I made! However it doesn't address importing my_tool from other packages. (i.e. if tools/ and other_pkg/ are siblings, I'd want to be able to import tools in other_pkg. Something I didn't really specify in my OP, though it was shown in the second diagram)

[CasualDemon]

Whoops, forgot the imports in __init__.py for that, fixed and PR updated.

[Louis]

I've edited my answer to include one way to address your concern.

[squanto773]

I think this is the correct way of doing this. I'd like to add how to install packages as editable: pip -e path/to/SomeProject link. This allows to keep your directory structure as you like, but you can still import the modules as any other module.

[CasualDemon]

@pod were you able to retest with the edits? Have any issues?

[Pod]

I've just tried them and they don't help, I'm afraid. Doing my_other_package/whatever.py results in import errors. However if I borrow from another answer and bodge the import path to use "." then it works.

[Pod]

I guess what I'm doing is a little bit insane, but this gets the closest and does so in ways that don't feel like I'm doing naughty things. Thanks :)

[Michael Beer]

That's hacky at best - don't mess around with os.path .